Plex: Towards Reliability using Pretrained Large Model Extensions